MCSBench - L2: leaderboard
Measures multimodal conceptual-structure alignment on MCSBench, a diagnostic multiple-choice benchmark of 661 base visual question-answering items over fine-grained visual and taxonomy-like categories. L2 is the relational inference level: inter-concept and multi-image relations. These are base-VQA accuracies from the released results table; the reasoning-chain integrity overlay, which carries the fourth metacognitive-verification level, is scored separately and is not part of this series. Five API models also have a run with web search enabled in that table; those runs are not part of this series either. This series reports MCSBench - L2 using Accuracy (%).
Metric: Accuracy (%). Source: huggingface.co. Saturation forecast: Around February 2027. 77 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3 Pro (Preview) | 84.65 | #71 |
| 2 | GPT-5 | 81.72 | #105 |
| 3 | Gemini 3 Flash (Preview) | 81.72 | #90 |
| 4 | Claude Opus 4.6 | 79.46 | #67 |
| 5 | Gemini 2.5 Pro | 79.01 | #159 |
| 6 | O4 Mini | 74.04 | #185 |
| 7 | Gemini 2.5 Flash | 73.14 | #255 |
| 8 | Qwen 3 VL 235B A22B (Thinking) | 71.56 | #236 |
| 9 | GPT-5 Mini | 71.33 | #193 |
| 10 | Qwen 3 VL 235B A22B Instruct | 68.62 | #280 |
| 11 | Qwen 3 VL 30B A3B Instruct | 66.59 | #390 |
| 12 | Qwen 2.5 VL 72B Instruct | 66.14 | #383 |
| 13 | Qwen 3 VL 30B A3B (Thinking) | 64.33 | #348 |
| 14 | Claude Sonnet 4.5 | 62.08 | #148 |
| 15 | Qwen 2.5 VL 32B Instruct | 61.4 | #463 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=mcsbench-l2 · How It Works · Data refreshed daily, snapshot 2026-11-05.